Neuron-Inspired Leader–Follower Networks for AI with Local Error Signals

Artificial intelligence (AI) systems often operate under memory and energy limits, similar to biological networks of neurons. The collective behavior of a network with heterogeneous, resource-limited information processing units (e.g., a group of fish, a flock of birds, or a network of neurons) demonstrates a high degree of self-organization, emergence, and complexity. These properties arise from simple interaction rules, where certain individuals can exhibit leadership-like behavior and influence the collective activity of the group. Motivated by collective intelligence in biological neural systems and neuron-level plasticity, we introduce a worker concept to an artificial neural network (NN) for AI. This NN structure contains workers that encompass one or more information processing units (e.g., neurons, filters, layers, or blocks of layers). Workers are either leaders or followers, and we train a leader-follower neural network (LFNN) by leveraging local error signals. The LFNN does not require backpropagation (BP) or a global loss function to achieve optimal performance (we denote LFNN trained without BP as LFNN-ℓ). By investigating worker behavior and evaluating the LFNN and LFNN-ℓ architectures on a variety of image classification tasks(e.g., MNIST, CIFAR-10, ImageNet), we demonstrate that LFNN-ℓ trained with local error signals achieves lower error rates and superior scalability than state-of-the-art machine learning approaches. Furthermore, LFNN-ℓ can be conveniently embedded in classic convolutional NN architectures (e.g., VGG, ResNet, and Vision Transformer (ViT)), achieving a 2x speedup compared to BP-based methods and significantly outperforming models trained with end-to-end BP and other state-of-the-art local learning methods in terms of accuracy on CIFAR-10, Tiny-ImageNet, and ImageNet. Lastly, the proposed LFNN-based model outperforms deep learning counterparts in brain-age prediction from magnetic resonance imaging (MRI) data, while also achieving a 2× speedup. These results indicate that neuron-inspired, decentralized training rules can improve AI accuracy and efficiency. The LFNN-ℓ architecture provides practical tools and insights for addressing core challenges in deep learning and for designing more biologically plausible and sustainable AI.